Voltage Optimization in Active Distribution Networks Based on Reinforcement Learning with a Voltage Safety Layer

The integration of Electric Vehicle (EV) and Distributed Generation (DG) introduces heightened uncertainties to distribution networks. To tackle the precision and real-time challenges inherent in conventional control methods employed for voltage control, this study proposes the Voltage Safety Layer-based Multi-Agent Deep Deterministic Policy Gradients (VSL-MADDPG) algorithm. Firstly, the active voltage optimization problem is reformulated as a Constrained Markov Decision Process (cMDP). Secondly, a Voltage Safety Layer based on the voltage sensitivity matrix is added to the MADDPG algorithm to mechanistically ensure the voltage security of the active distribution network. Finally, results of case simulations on the IEEE33 node system demonstrate that the proposed algorithm significantly reduces the network losses and the charging costs.